es-ude/OnDeviceTraining

Framework for Inference and Training of Deep Neural Networks on MCU + PC

Stars: 1
Forks: 3
Open beginner issues: 0
Indexed issues: 76
Avg merge: 1d 1h
Merged PRs (30d): 8

Dominant language: C
License: MIT
Last GitHub push: Jul 21, 2026
Latest indexed: Sep 15, 2026
Contributing guide: Contributing guide
Code of conduct: No code of conduct
Beginner labels: No beginner labels indexed

Open All

76 open issues indexed

Implement Gradient Checkpointing Open

enhancement

es-ude/OnDeviceTraining#4 · 0 comments · 0 reactions · 1 assignee · Indexed Sep 15, 2026
Change functions from pointers to handles Open

enhancement

es-ude/OnDeviceTraining#15 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Implement Conv2d Open

es-ude/OnDeviceTraining#47 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Epic: ODT Python Pipeline — Providers, ir2c, Resource Estimation Open

enhancement

es-ude/OnDeviceTraining#58 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Implement TrainingImplementationProviders (Float32, SymInt32, Asym) Open

enhancement

es-ude/OnDeviceTraining#60 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Implement InferenceImplementationProviders (Float32, SymInt32, Asym) Open

enhancement

es-ude/OnDeviceTraining#61 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Implement ir2c — C code generation from annotated IR Open

enhancement

es-ude/OnDeviceTraining#62 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Implement ResourceEstimator (cycles, memory, flash) Open

enhancement

es-ude/OnDeviceTraining#63 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
End-to-end integration tests (PyTorch → ir2c → compile → run) Open

enhancement

es-ude/OnDeviceTraining#64 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
userApi: add ordered cleanup primitive for model + optimizer + auxiliary tensors Open

es-ude/OnDeviceTraining#113 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Implement fully quantized training Open

es-ude/OnDeviceTraining#137 · 1 comment · 0 reactions · 1 assignee · Indexed Sep 15, 2026
static Open

es-ude/OnDeviceTraining#138 · 0 comments · 0 reactions · 1 assignee · Indexed Sep 15, 2026
dynamic Open

es-ude/OnDeviceTraining#139 · 0 comments · 0 reactions · 1 assignee · Indexed Sep 15, 2026
1. Quantize some layers during forward pass Open

es-ude/OnDeviceTraining#140 · 0 comments · 0 reactions · 1 assignee · Indexed Sep 15, 2026
2. Quantize all layers during forward pass Open

es-ude/OnDeviceTraining#141 · 0 comments · 0 reactions · 1 assignee · Indexed Sep 15, 2026
3. Quantize all layers during forward and backward pass Open

es-ude/OnDeviceTraining#142 · 0 comments · 0 reactions · 1 assignee · Indexed Sep 15, 2026
Add weighted loss variants (MSE_WEIGHTED, CROSS_ENTROPY_WEIGHTED) Open

es-ude/OnDeviceTraining#150 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Common.h: add PRINT_WARN + unified assert macros Open

es-ude/OnDeviceTraining#151 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Use microbatch B>1 not possible Open

es-ude/OnDeviceTraining#152 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Microbatch shape contract asserts in loss forward/backward Open

es-ude/OnDeviceTraining#153 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
tensor: N=0 element allocation has implementation-defined behavior Open

es-ude/OnDeviceTraining#160 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
examples: full-SYM ECG reconstruction AE (Conv SYM + Quantization layers + MSE) Open

enhancement

es-ude/OnDeviceTraining#207 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 15, 2026
test: shared goldgen Python module + C tensor-builder helper lib + remove dead MORE_SOURCES Open

enhancement

es-ude/OnDeviceTraining#208 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
docs: mutation-class catalog, padding-convention note, integer-only requant open problem, SYM pool mechanics Open

documentation

es-ude/OnDeviceTraining#209 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
SYM completeness: requant primitives + Quantization layer + layer SYM arms Open

enhancement

es-ude/OnDeviceTraining#210 · 2 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
research: quantized gradient-accumulation scheme — fixed-scale vs strategy A, and a Deutel attribution correction Open

enhancement

es-ude/OnDeviceTraining#218 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
build: consolidate circular micro static-libraries to remove link-line duplication at the source Open

es-ude/OnDeviceTraining#238 · 2 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
examples/kws: replace module-level assert KWS_CLASSES validation with explicit raise Open

es-ude/OnDeviceTraining#258 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Dropout factory gap: missing Owning and layerQuant_t variants Open

es-ude/OnDeviceTraining#268 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
LayerNorm float backward: route dgamma/dbeta through the identity-kernel funnel (packed LN grads) Open

es-ude/OnDeviceTraining#275 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
research: C-vs-PyTorch training-dynamics divergence (bit-parity-invisible) Open

research

es-ude/OnDeviceTraining#299 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
har_classifier: map the quantization design space (epic) Open

enhancement research

es-ude/OnDeviceTraining#300 · 5 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
optimizer: rename optimizer_t.sizeStates — it counts tracked parameters, not states Open

es-ude/OnDeviceTraining#314 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
userApi: wire allocators hard-exit on packed SYM/ASYM — packed activation/dx wires structurally impossible Open

enhancement

es-ude/OnDeviceTraining#317 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
training_loop: loss-grad seed wire has no storage-quantization knob — inherits model-output dtype Open

enhancement

es-ude/OnDeviceTraining#318 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
training_loop: non-sequential model topology — branch/merge DAG for forward+backward (epic) Open

enhancement

es-ude/OnDeviceTraining#329 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
layer/arithmetic: 2D sliding-window layer family for image recognition (epic) Open

enhancement

es-ude/OnDeviceTraining#330 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
layer: ResNet-style skip connections for 1D and 2D CNNs Open

enhancement

es-ude/OnDeviceTraining#331 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
layer: DenseNet-style dense blocks for 1D and 2D CNNs Open

enhancement

es-ude/OnDeviceTraining#332 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
model: small time-series transformer architectures Open

enhancement

es-ude/OnDeviceTraining#333 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
arithmetic: flex-point / block-floating-point tensor format for FPU-less MCUs Open

enhancement research

es-ude/OnDeviceTraining#335 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
arithmetic: evaluate third-party kernel libraries (CMSIS-NN/DSP, …) vs hand-rolled kernels Open

research

es-ude/OnDeviceTraining#336 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
continual-learning: ARITH_SYM_INT32 arms for PPCA ops (integer eigensolver, literature-first) Open

enhancement research

es-ude/OnDeviceTraining#343 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
data_loader: transform/targetTransform fields are declared but never invoked Open

bug

es-ude/OnDeviceTraining#345 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
continual-learning: post-#326 hardening follow-ups (test gaps, CCIPCA in-place, CI gate order) Open

enhancement

es-ude/OnDeviceTraining#346 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
build: FP contraction diverges host vs MCU — decide global -ffp-contract policy Open

enhancement

es-ude/OnDeviceTraining#349 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
training-state checkpointing: optimizer/scheduler/RNG state is not serializable — resume silently corrupts training Open

enhancement

es-ude/OnDeviceTraining#350 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
arithmetic: enable TRACK_INSTRUCTIONS on the legacy op libs Open

es-ude/OnDeviceTraining#351 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
tensor: ASYM int-repr dequant codes[i] + zeroPoint can overflow int32 for far-from-zero positive grids Open

es-ude/OnDeviceTraining#384 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
arithmetic/executeOp: writeOut swap-restore of the target's qConfig roundingMode is not reentrant — replace with explicit rounding parameter when multi-threading lands Open

es-ude/OnDeviceTraining#386 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026

A short digest of beginner-friendly GitHub issues.